def layer_op(self, input_tensor, is_training=None, keep_prob=None):
channel_dim = len(input_tensor.get_shape()) - 1
stack = [input_tensor]
input_mask = tf.ones([input_tensor.get_shape().as_list()[-1]]) > 0
for idx, d in enumerate(self.dilation_rates):
if idx == len(self.dilation_rates) - 1:
keep_prob = None # no dropout on last layer of the stack
if self.use_bdo:
conv = ChannelSparseConvolutionalLayer(
self.n_dense_channels,
kernel_size=self.kernel_size,
**self.kwargs)
conv, new_input_mask = conv(tf.concat(stack, channel_dim),
input_mask=input_mask,
is_training=is_training,
keep_prob=keep_prob)
input_mask = tf.concat([input_mask, new_input_mask], 0)
else:
conv = ConvolutionalLayer(self.n_dense_channels,
kernel_size=self.kernel_size,
**self.kwargs)
conv = conv(tf.concat(stack, channel_dim),
is_training=is_training,
keep_prob=keep_prob)
stack.append(conv)
return stack
def test_masks(self):
x = tf.random_normal(shape=[2,4,5,4])
conv1 = ChannelSparseConvolutionalLayer(10)
conv2 = ChannelSparseConvolutionalLayer(10)
conv3 = ChannelSparseConvolutionalLayer(10)
conv4 = ChannelSparseConvolutionalLayer(10)
x1, mask1=conv1(x, None, True, 1.)
x2, mask2=conv2(x1, mask1, True, .5)
x3, mask3=conv3(x2, mask2, True, .2)
x4, mask4=conv4(x3, mask3, True, 1.)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
out1, out2, out3, out4 = sess.run([mask1, mask2, mask3, mask4])
self.assertAllClose([10, 5, 2, 10], [np.sum(out1),
np.sum(out2),
np.sum(out3),
np.sum(out4)])
def test_2d_shape(self):
x = tf.random_normal(shape=[2, 4, 5, 4])
conv1 = ChannelSparseConvolutionalLayer(4)
conv2 = ChannelSparseConvolutionalLayer(8, kernel_size=[1, 1, 3])
conv3 = ChannelSparseConvolutionalLayer(4, acti_func='relu')
conv4 = ChannelSparseConvolutionalLayer(8, with_bn=False)
conv5 = ChannelSparseConvolutionalLayer(4, with_bias=True)
x1, mask1 = conv1(x, None, True, 1.)
x2, mask2 = conv2(x1, mask1, True, 1.)
x3, mask3 = conv3(x2, mask2, True, .5)
x4, mask4 = conv4(x3, mask3, True, .75)
x5, mask5 = conv5(x4, mask4, True, 1.)
with self.test_session(config=get_config()) as sess:
sess.run(tf.global_variables_initializer())
out1, out2, out3, out4, out5 = sess.run([x1, x2, x3, x4, x5])
self.assertAllClose([2, 4, 5, 4], out1.shape)
self.assertAllClose([2, 4, 5, 8], out2.shape)
self.assertAllClose([2, 4, 5, 2], out3.shape)
self.assertAllClose([2, 4, 5, 6], out4.shape)
self.assertAllClose([2, 4, 5, 4], out5.shape)
def create_block(self):
dfs_block = []
for _ in self.dilation_rates:
if self.use_bdo:
conv = ChannelSparseConvolutionalLayer(
self.n_dense_channels,
kernel_size=self.kernel_size,
**self.kwargs)
else:
conv = ConvolutionalLayer(self.n_dense_channels,
kernel_size=self.kernel_size,
**self.kwargs)
dfs_block.append(conv)
return dfs_block
def create_block(self):
net_conv_layers = []
for _ in self.dilation_rates:
if self.use_bdo:
conv = ChannelSparseConvolutionalLayer(
self.n_dense_channels,
kernel_size=self.kernel_size,
**self.kwargs)
else:
conv = ConvolutionalLayer(self.n_dense_channels,
kernel_size=self.kernel_size,
**self.kwargs)
net_conv_layers.append(conv)
return DenseFSBlockDesc(conv_layers=net_conv_layers)